The paper identifies and describes promising architectural solutions for convolutional neural networks, along with key parameters for further structural and parametric synthesis. The advantages and disadvantages of different blocks are outlined, and their relevance in structural synthesis is substantiated. The use of a genetic evolutionary algorithm for structural-parametric synthesis is proposed, along with a review of contemporary approaches. The configuration process of the evolutionary algorithm is shown and described. Based on optimization criteria, the fitness function, selection, mutation, and crossover methods are defined. The results of the experimental evolutionary process are presented and analyzed. An example model created by evolutionary algorithms is considered, which is based on functional blocks aggregated from various architectural approaches in convolutional neural networks. For each model, performance criteria, including average reduction in training time, benefits, and architectural integration details, were calculated during the synthesis process. Experimental results demonstrate that using complex structural blocks instead of traditional layers with a flexible fitness function configuration according to quality and performance criteria yields significant improvements for the final model
Read full abstract7-days of FREE Audio papers, translation & more with Prime
7-days of FREE Prime access